high performance computing on graphics processing units: hgpu.org

The paper addresses the problem of pose-invariant recognition of faces via an MRF matching model. Unlike previous costly matching approaches, the proposed algorithm employs effective techniques to reduce the MRF inference time. To this end, processing is done in a parallel fashion on a GPU employing a dual decomposition framework. The optimisation is further accelerated taking a multi-resolution approach based on the Renormalisation Group Theory (RGT) along with efficient methods for message passing and the incremental subgradient approach. For the graph construction, Daisy features are used as node attributes exhibiting high cross-pose invariance, while high discriminatory capability in the classification stage is obtained via multi-scale LBP histograms. The experimental evaluation of the method is performed via extensive tests on the databases of XM2VTS, FERET and LFW in verification, identification and the unseen pair-matching paradigms. The proposed approach achieves state-of-the-art performance in pose-invariant recognition of faces and performs as well or better than the existing methods in the unconstrained settings of the challenging LFW database using a single feature for classification.